Survival analysis with censored outcomes

Abstract

This dissertation is arranged in the three-paper format. The unifying theme of the dissertation is estimating longevity in terms of years of life lost (YLL). The papers are presented in order of increasing methodological focus.

The first paper, No compelling evidence that sibutramine prolongs life in rodents despite providing a dose-dependent reduction in body weight, was published in the International Journal of Obesity in 2010. Although sibutramine reduced food intake and body weight, it had no association with longevity. The conclusions were based on data from a 2-year toxicology trial required by the FDA. Although this paper was largely applied research, we did demonstrate that data truncated at two years can be re-purposed for drawing conclusions about total longevity; potentially, there exist many other data sets that can be utilized for inferring similar conclusions about total longevity.

The second paper, Can rodent longevity studies be both short and powerful? was published in the Journal of Gerontology: Biological Sciences in 2010. This paper was more methodological; we amassed data from 15 rodent longevity studies carried through to completion, to assess whether a larger study conducted over a shorter interval could have reached the same conclusions concerning significant differences in longevity. We concluded that as long as 20% of the rodents die in a given time period, a 5-fold increase in sample size will have equivalent power to a full-length study.

The third paper, the capstone of my dissertation, is Turning the Analysis of Risk Factor Mortality Associations Upside Down: Modeling Conditional Expectations, Years of Life Lost Through a Novel Family of Generalized Normal Distributions. Here, I derived a new family of generalized normal distributions and demonstrate that a member of this family is able to model human longevity more accurately than existing alternatives (Cox, Weibull, and Gompertz models). Keywords: Survival analysis, generalized normal, demographics, parametric models, YLL.